2011-04-04 13:30:58 +08:00
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
|
2011-04-04 13:30:58 +08:00
|
|
|
*
|
|
|
|
* Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
|
2012-11-23 19:03:19 +08:00
|
|
|
* Copyright (c) 2012 Paolo Valente.
|
2011-04-04 13:30:58 +08:00
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public License
|
|
|
|
* version 2 as published by the Free Software Foundation.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/module.h>
|
|
|
|
#include <linux/init.h>
|
|
|
|
#include <linux/bitops.h>
|
|
|
|
#include <linux/errno.h>
|
|
|
|
#include <linux/netdevice.h>
|
|
|
|
#include <linux/pkt_sched.h>
|
|
|
|
#include <net/sch_generic.h>
|
|
|
|
#include <net/pkt_sched.h>
|
|
|
|
#include <net/pkt_cls.h>
|
|
|
|
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Quick Fair Queueing Plus
|
|
|
|
========================
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
Sources:
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
[1] Paolo Valente,
|
|
|
|
"Reducing the Execution Time of Fair-Queueing Schedulers."
|
|
|
|
http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
|
|
|
|
|
|
|
|
Sources for QFQ:
|
|
|
|
|
|
|
|
[2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
|
2011-04-04 13:30:58 +08:00
|
|
|
Packet Scheduling with Tight Bandwidth Distribution Guarantees."
|
|
|
|
|
|
|
|
See also:
|
|
|
|
http://retis.sssup.it/~fabio/linux/qfq/
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
|
|
|
|
classes. Each aggregate is timestamped with a virtual start time S
|
|
|
|
and a virtual finish time F, and scheduled according to its
|
|
|
|
timestamps. S and F are computed as a function of a system virtual
|
|
|
|
time function V. The classes within each aggregate are instead
|
|
|
|
scheduled with DRR.
|
|
|
|
|
|
|
|
To speed up operations, QFQ+ divides also aggregates into a limited
|
|
|
|
number of groups. Which group a class belongs to depends on the
|
|
|
|
ratio between the maximum packet length for the class and the weight
|
|
|
|
of the class. Groups have their own S and F. In the end, QFQ+
|
|
|
|
schedules groups, then aggregates within groups, then classes within
|
|
|
|
aggregates. See [1] and [2] for a full description.
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
Virtual time computations.
|
|
|
|
|
|
|
|
S, F and V are all computed in fixed point arithmetic with
|
|
|
|
FRAC_BITS decimal bits.
|
|
|
|
|
|
|
|
QFQ_MAX_INDEX is the maximum index allowed for a group. We need
|
|
|
|
one bit per index.
|
|
|
|
QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
|
|
|
|
|
|
|
|
The layout of the bits is as below:
|
|
|
|
|
|
|
|
[ MTU_SHIFT ][ FRAC_BITS ]
|
|
|
|
[ MAX_INDEX ][ MIN_SLOT_SHIFT ]
|
|
|
|
^.__grp->index = 0
|
|
|
|
*.__grp->slot_shift
|
|
|
|
|
|
|
|
where MIN_SLOT_SHIFT is derived by difference from the others.
|
|
|
|
|
|
|
|
The max group index corresponds to Lmax/w_min, where
|
|
|
|
Lmax=1<<MTU_SHIFT, w_min = 1 .
|
|
|
|
From this, and knowing how many groups (MAX_INDEX) we want,
|
|
|
|
we can derive the shift corresponding to each group.
|
|
|
|
|
|
|
|
Because we often need to compute
|
|
|
|
F = S + len/w_i and V = V + len/wsum
|
|
|
|
instead of storing w_i store the value
|
|
|
|
inv_w = (1<<FRAC_BITS)/w_i
|
|
|
|
so we can do F = S + len * inv_w * wsum.
|
|
|
|
We use W_TOT in the formulas so we can easily move between
|
|
|
|
static and adaptive weight sum.
|
|
|
|
|
|
|
|
The per-scheduler-instance data contain all the data structures
|
|
|
|
for the scheduler: bitmaps and bucket lists.
|
|
|
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Maximum number of consecutive slots occupied by backlogged classes
|
|
|
|
* inside a group.
|
|
|
|
*/
|
|
|
|
#define QFQ_MAX_SLOTS 32
|
|
|
|
|
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Shifts used for aggregate<->group mapping. We allow class weights that are
|
|
|
|
* in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
|
2011-04-04 13:30:58 +08:00
|
|
|
* group with the smallest index that can support the L_i / r_i configured
|
2012-11-23 19:03:19 +08:00
|
|
|
* for the classes in the aggregate.
|
2011-04-04 13:30:58 +08:00
|
|
|
*
|
|
|
|
* grp->index is the index of the group; and grp->slot_shift
|
|
|
|
* is the shift for the corresponding (scaled) sigma_i.
|
|
|
|
*/
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
#define QFQ_MAX_INDEX 24
|
2012-11-23 19:03:19 +08:00
|
|
|
#define QFQ_MAX_WSHIFT 10
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
|
|
|
|
#define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
#define FRAC_BITS 30 /* fixed point arithmetic */
|
|
|
|
#define ONE_FP (1UL << FRAC_BITS)
|
|
|
|
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
#define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
|
2012-11-23 19:03:19 +08:00
|
|
|
#define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
|
|
|
|
|
|
|
|
#define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Possible group states. These values are used as indexes for the bitmaps
|
|
|
|
* array of struct qfq_queue.
|
|
|
|
*/
|
|
|
|
enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
|
|
|
|
|
|
|
|
struct qfq_group;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate;
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
struct qfq_class {
|
|
|
|
struct Qdisc_class_common common;
|
|
|
|
|
|
|
|
unsigned int filter_cnt;
|
|
|
|
|
|
|
|
struct gnet_stats_basic_packed bstats;
|
|
|
|
struct gnet_stats_queue qstats;
|
2016-12-05 01:48:16 +08:00
|
|
|
struct net_rate_estimator __rcu *rate_est;
|
2011-04-04 13:30:58 +08:00
|
|
|
struct Qdisc *qdisc;
|
2012-11-23 19:03:19 +08:00
|
|
|
struct list_head alist; /* Link for active-classes list. */
|
|
|
|
struct qfq_aggregate *agg; /* Parent aggregate. */
|
|
|
|
int deficit; /* DRR deficit counter. */
|
|
|
|
};
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate {
|
2011-04-04 13:30:58 +08:00
|
|
|
struct hlist_node next; /* Link for the slot list. */
|
|
|
|
u64 S, F; /* flow timestamps (exact) */
|
|
|
|
|
|
|
|
/* group we belong to. In principle we would need the index,
|
|
|
|
* which is log_2(lmax/weight), but we never reference it
|
|
|
|
* directly, only the group.
|
|
|
|
*/
|
|
|
|
struct qfq_group *grp;
|
|
|
|
|
|
|
|
/* these are copied from the flowset. */
|
2012-11-23 19:03:19 +08:00
|
|
|
u32 class_weight; /* Weight of each class in this aggregate. */
|
|
|
|
/* Max pkt size for the classes in this aggregate, DRR quantum. */
|
|
|
|
int lmax;
|
|
|
|
|
|
|
|
u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
|
|
|
|
u32 budgetmax; /* Max budget for this aggregate. */
|
|
|
|
u32 initial_budget, budget; /* Initial and current budget. */
|
|
|
|
|
|
|
|
int num_classes; /* Number of classes in this aggr. */
|
|
|
|
struct list_head active; /* DRR queue of active classes. */
|
|
|
|
|
|
|
|
struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
|
2011-04-04 13:30:58 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
struct qfq_group {
|
|
|
|
u64 S, F; /* group timestamps (approx). */
|
|
|
|
unsigned int slot_shift; /* Slot shift. */
|
|
|
|
unsigned int index; /* Group index. */
|
|
|
|
unsigned int front; /* Index of the front slot. */
|
|
|
|
unsigned long full_slots; /* non-empty slots */
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Array of RR lists of active aggregates. */
|
2011-04-04 13:30:58 +08:00
|
|
|
struct hlist_head slots[QFQ_MAX_SLOTS];
|
|
|
|
};
|
|
|
|
|
|
|
|
struct qfq_sched {
|
2014-09-13 11:05:27 +08:00
|
|
|
struct tcf_proto __rcu *filter_list;
|
2017-05-17 17:07:55 +08:00
|
|
|
struct tcf_block *block;
|
2011-04-04 13:30:58 +08:00
|
|
|
struct Qdisc_class_hash clhash;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
u64 oldV, V; /* Precise virtual times. */
|
|
|
|
struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
|
|
|
|
u32 wsum; /* weight sum */
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
u32 iwsum; /* inverse weight sum */
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
|
|
|
|
struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
|
2012-11-23 19:03:19 +08:00
|
|
|
u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
|
|
|
|
|
|
|
|
u32 max_agg_classes; /* Max number of classes per aggr. */
|
|
|
|
struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
|
2011-04-04 13:30:58 +08:00
|
|
|
};
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/*
|
|
|
|
* Possible reasons why the timestamps of an aggregate are updated
|
|
|
|
* enqueue: the aggregate switches from idle to active and must scheduled
|
|
|
|
* for service
|
|
|
|
* requeue: the aggregate finishes its budget, so it stops being served and
|
|
|
|
* must be rescheduled for service
|
|
|
|
*/
|
|
|
|
enum update_reason {enqueue, requeue};
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct Qdisc_class_common *clc;
|
|
|
|
|
|
|
|
clc = qdisc_class_find(&q->clhash, classid);
|
|
|
|
if (clc == NULL)
|
|
|
|
return NULL;
|
|
|
|
return container_of(clc, struct qfq_class, common);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_purge_queue(struct qfq_class *cl)
|
|
|
|
{
|
|
|
|
unsigned int len = cl->qdisc->q.qlen;
|
2016-02-26 06:55:01 +08:00
|
|
|
unsigned int backlog = cl->qdisc->qstats.backlog;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
qdisc_reset(cl->qdisc);
|
2016-02-26 06:55:01 +08:00
|
|
|
qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
|
|
|
|
[TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
|
|
|
|
[TCA_QFQ_LMAX] = { .type = NLA_U32 },
|
|
|
|
};
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Calculate a flow index, given its weight and maximum packet length.
|
|
|
|
* index = log_2(maxlen/weight) but we need to apply the scaling.
|
|
|
|
* This is used only once at flow creation.
|
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
u64 slot_size = (u64)maxlen * inv_w;
|
|
|
|
unsigned long size_map;
|
|
|
|
int index = 0;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
size_map = slot_size >> min_slot_shift;
|
2011-04-04 13:30:58 +08:00
|
|
|
if (!size_map)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
index = __fls(size_map) + 1; /* basically a log_2 */
|
2012-11-23 19:03:19 +08:00
|
|
|
index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
if (index < 0)
|
|
|
|
index = 0;
|
|
|
|
out:
|
|
|
|
pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
|
|
|
|
(unsigned long) ONE_FP/inv_w, maxlen, index);
|
|
|
|
|
|
|
|
return index;
|
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
|
|
|
|
static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
|
|
|
|
enum update_reason);
|
|
|
|
|
|
|
|
static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
|
|
|
|
u32 lmax, u32 weight)
|
2012-08-07 15:27:25 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
INIT_LIST_HEAD(&agg->active);
|
|
|
|
hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
|
|
|
|
|
|
|
|
agg->lmax = lmax;
|
|
|
|
agg->class_weight = weight;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
|
|
|
|
u32 lmax, u32 weight)
|
|
|
|
{
|
|
|
|
struct qfq_aggregate *agg;
|
|
|
|
|
hlist: drop the node parameter from iterators
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 09:06:00 +08:00
|
|
|
hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
|
2012-11-23 19:03:19 +08:00
|
|
|
if (agg->lmax == lmax && agg->class_weight == weight)
|
|
|
|
return agg;
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2012-08-07 15:27:25 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Update aggregate as a function of the new number of classes. */
|
|
|
|
static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
|
|
|
|
int new_num_classes)
|
|
|
|
{
|
|
|
|
u32 new_agg_weight;
|
|
|
|
|
|
|
|
if (new_num_classes == q->max_agg_classes)
|
|
|
|
hlist_del_init(&agg->nonfull_next);
|
|
|
|
|
|
|
|
if (agg->num_classes > new_num_classes &&
|
|
|
|
new_num_classes == q->max_agg_classes - 1) /* agg no more full */
|
|
|
|
hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
|
|
|
|
|
2013-03-05 16:04:57 +08:00
|
|
|
/* The next assignment may let
|
|
|
|
* agg->initial_budget > agg->budgetmax
|
|
|
|
* hold, we will take it into account in charge_actual_service().
|
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
agg->budgetmax = new_num_classes * agg->lmax;
|
|
|
|
new_agg_weight = agg->class_weight * new_num_classes;
|
|
|
|
agg->inv_w = ONE_FP/new_agg_weight;
|
|
|
|
|
|
|
|
if (agg->grp == NULL) {
|
|
|
|
int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
|
|
|
|
q->min_slot_shift);
|
|
|
|
agg->grp = &q->groups[i];
|
|
|
|
}
|
|
|
|
|
|
|
|
q->wsum +=
|
|
|
|
(int) agg->class_weight * (new_num_classes - agg->num_classes);
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
q->iwsum = ONE_FP / q->wsum;
|
2012-11-23 19:03:19 +08:00
|
|
|
|
|
|
|
agg->num_classes = new_num_classes;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Add class to aggregate. */
|
|
|
|
static void qfq_add_to_agg(struct qfq_sched *q,
|
|
|
|
struct qfq_aggregate *agg,
|
|
|
|
struct qfq_class *cl)
|
|
|
|
{
|
|
|
|
cl->agg = agg;
|
|
|
|
|
|
|
|
qfq_update_agg(q, agg, agg->num_classes+1);
|
|
|
|
if (cl->qdisc->q.qlen > 0) { /* adding an active class */
|
|
|
|
list_add_tail(&cl->alist, &agg->active);
|
|
|
|
if (list_first_entry(&agg->active, struct qfq_class, alist) ==
|
|
|
|
cl && q->in_serv_agg != agg) /* agg was inactive */
|
|
|
|
qfq_activate_agg(q, agg, enqueue); /* schedule agg */
|
|
|
|
}
|
2012-08-07 15:27:25 +08:00
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
|
2012-08-07 15:27:25 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
|
2012-08-07 15:27:25 +08:00
|
|
|
{
|
2015-06-17 06:16:59 +08:00
|
|
|
hlist_del_init(&agg->nonfull_next);
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
q->wsum -= agg->class_weight;
|
|
|
|
if (q->wsum != 0)
|
|
|
|
q->iwsum = ONE_FP / q->wsum;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (q->in_serv_agg == agg)
|
|
|
|
q->in_serv_agg = qfq_choose_next_agg(q);
|
|
|
|
kfree(agg);
|
|
|
|
}
|
2012-08-07 15:27:25 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Deschedule class from within its parent aggregate. */
|
|
|
|
static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
|
|
|
|
{
|
|
|
|
struct qfq_aggregate *agg = cl->agg;
|
2012-08-07 15:27:25 +08:00
|
|
|
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
list_del(&cl->alist); /* remove from RR queue of the aggregate */
|
|
|
|
if (list_empty(&agg->active)) /* agg is now inactive */
|
|
|
|
qfq_deactivate_agg(q, agg);
|
2012-08-07 15:27:25 +08:00
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Remove class from its parent aggregate. */
|
|
|
|
static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *agg = cl->agg;
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
cl->agg = NULL;
|
|
|
|
if (agg->num_classes == 1) { /* agg being emptied, destroy it */
|
|
|
|
qfq_destroy_agg(q, agg);
|
|
|
|
return;
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
}
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_update_agg(q, agg, agg->num_classes-1);
|
|
|
|
}
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Deschedule class and remove it from its parent aggregate. */
|
|
|
|
static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
|
|
|
|
{
|
|
|
|
if (cl->qdisc->q.qlen > 0) /* class is active */
|
|
|
|
qfq_deactivate_class(q, cl);
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_rm_from_agg(q, cl);
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Move class to a new aggregate, matching the new class weight and/or lmax */
|
|
|
|
static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
|
|
|
|
u32 lmax)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
|
|
|
|
|
|
|
|
if (new_agg == NULL) { /* create new aggregate */
|
|
|
|
new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
|
|
|
|
if (new_agg == NULL)
|
|
|
|
return -ENOBUFS;
|
|
|
|
qfq_init_agg(q, new_agg, lmax, weight);
|
|
|
|
}
|
|
|
|
qfq_deact_rm_from_agg(q, cl);
|
|
|
|
qfq_add_to_agg(q, new_agg, cl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
|
2017-12-21 01:35:15 +08:00
|
|
|
struct nlattr **tca, unsigned long *arg,
|
|
|
|
struct netlink_ext_ack *extack)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)*arg;
|
2012-11-23 19:03:19 +08:00
|
|
|
bool existing = false;
|
2011-04-04 13:30:58 +08:00
|
|
|
struct nlattr *tb[TCA_QFQ_MAX + 1];
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *new_agg = NULL;
|
2011-04-04 13:30:58 +08:00
|
|
|
u32 weight, lmax, inv_w;
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
int err;
|
2012-01-02 19:47:50 +08:00
|
|
|
int delta_w;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
if (tca[TCA_OPTIONS] == NULL) {
|
|
|
|
pr_notice("qfq: no options\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2017-04-12 20:34:07 +08:00
|
|
|
err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy,
|
|
|
|
NULL);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (err < 0)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
if (tb[TCA_QFQ_WEIGHT]) {
|
|
|
|
weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
|
|
|
|
if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
|
|
|
|
pr_notice("qfq: invalid weight %u\n", weight);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
} else
|
|
|
|
weight = 1;
|
|
|
|
|
|
|
|
if (tb[TCA_QFQ_LMAX]) {
|
|
|
|
lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
|
2011-04-04 13:30:58 +08:00
|
|
|
pr_notice("qfq: invalid max length %u\n", lmax);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
} else
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
lmax = psched_mtu(qdisc_dev(sch));
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
inv_w = ONE_FP / weight;
|
|
|
|
weight = ONE_FP / inv_w;
|
|
|
|
|
|
|
|
if (cl != NULL &&
|
|
|
|
lmax == cl->agg->lmax &&
|
|
|
|
weight == cl->agg->class_weight)
|
|
|
|
return 0; /* nothing to change */
|
|
|
|
|
|
|
|
delta_w = weight - (cl ? cl->agg->class_weight : 0);
|
|
|
|
|
|
|
|
if (q->wsum + delta_w > QFQ_MAX_WSUM) {
|
|
|
|
pr_notice("qfq: total weight out of range (%d + %u)\n",
|
|
|
|
delta_w, q->wsum);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (cl != NULL) { /* modify existing class */
|
2011-04-04 13:30:58 +08:00
|
|
|
if (tca[TCA_RATE]) {
|
2014-09-29 02:52:56 +08:00
|
|
|
err = gen_replace_estimator(&cl->bstats, NULL,
|
|
|
|
&cl->rate_est,
|
2016-06-07 00:37:16 +08:00
|
|
|
NULL,
|
|
|
|
qdisc_root_sleeping_running(sch),
|
2011-04-04 13:30:58 +08:00
|
|
|
tca[TCA_RATE]);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
}
|
2012-11-23 19:03:19 +08:00
|
|
|
existing = true;
|
|
|
|
goto set_change_agg;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* create and init new class */
|
2011-04-04 13:30:58 +08:00
|
|
|
cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
|
|
|
|
if (cl == NULL)
|
|
|
|
return -ENOBUFS;
|
|
|
|
|
|
|
|
cl->common.classid = classid;
|
2012-11-23 19:03:19 +08:00
|
|
|
cl->deficit = lmax;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2017-12-21 01:35:21 +08:00
|
|
|
cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
|
|
|
|
classid, NULL);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (cl->qdisc == NULL)
|
|
|
|
cl->qdisc = &noop_qdisc;
|
|
|
|
|
|
|
|
if (tca[TCA_RATE]) {
|
2014-09-29 02:52:56 +08:00
|
|
|
err = gen_new_estimator(&cl->bstats, NULL,
|
|
|
|
&cl->rate_est,
|
2016-06-07 00:37:16 +08:00
|
|
|
NULL,
|
|
|
|
qdisc_root_sleeping_running(sch),
|
2011-04-04 13:30:58 +08:00
|
|
|
tca[TCA_RATE]);
|
2012-11-23 19:03:19 +08:00
|
|
|
if (err)
|
|
|
|
goto destroy_class;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2017-03-08 23:03:32 +08:00
|
|
|
if (cl->qdisc != &noop_qdisc)
|
|
|
|
qdisc_hash_add(cl->qdisc, true);
|
2011-04-04 13:30:58 +08:00
|
|
|
sch_tree_lock(sch);
|
|
|
|
qdisc_class_hash_insert(&q->clhash, &cl->common);
|
|
|
|
sch_tree_unlock(sch);
|
|
|
|
|
|
|
|
qdisc_class_hash_grow(sch, &q->clhash);
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
set_change_agg:
|
|
|
|
sch_tree_lock(sch);
|
|
|
|
new_agg = qfq_find_agg(q, lmax, weight);
|
|
|
|
if (new_agg == NULL) { /* create new aggregate */
|
|
|
|
sch_tree_unlock(sch);
|
|
|
|
new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
|
|
|
|
if (new_agg == NULL) {
|
|
|
|
err = -ENOBUFS;
|
2016-12-05 01:48:16 +08:00
|
|
|
gen_kill_estimator(&cl->rate_est);
|
2012-11-23 19:03:19 +08:00
|
|
|
goto destroy_class;
|
|
|
|
}
|
|
|
|
sch_tree_lock(sch);
|
|
|
|
qfq_init_agg(q, new_agg, lmax, weight);
|
|
|
|
}
|
|
|
|
if (existing)
|
|
|
|
qfq_deact_rm_from_agg(q, cl);
|
|
|
|
qfq_add_to_agg(q, new_agg, cl);
|
|
|
|
sch_tree_unlock(sch);
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
*arg = (unsigned long)cl;
|
|
|
|
return 0;
|
2012-11-23 19:03:19 +08:00
|
|
|
|
|
|
|
destroy_class:
|
2018-09-25 00:22:50 +08:00
|
|
|
qdisc_put(cl->qdisc);
|
2012-11-23 19:03:19 +08:00
|
|
|
kfree(cl);
|
|
|
|
return err;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_rm_from_agg(q, cl);
|
2016-12-05 01:48:16 +08:00
|
|
|
gen_kill_estimator(&cl->rate_est);
|
2018-09-25 00:22:50 +08:00
|
|
|
qdisc_put(cl->qdisc);
|
2011-04-04 13:30:58 +08:00
|
|
|
kfree(cl);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
|
|
|
|
if (cl->filter_cnt > 0)
|
|
|
|
return -EBUSY;
|
|
|
|
|
|
|
|
sch_tree_lock(sch);
|
|
|
|
|
|
|
|
qfq_purge_queue(cl);
|
|
|
|
qdisc_class_hash_remove(&q->clhash, &cl->common);
|
|
|
|
|
|
|
|
sch_tree_unlock(sch);
|
|
|
|
|
net_sched: remove tc class reference counting
For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:
1) For class modification and dumping paths, we already hold RTNL lock,
so all of these ->get(),->change(),->put() are atomic.
2) For filter bindiing/unbinding, we use other reference counter than
this one, and they should have RTNL lock too.
3) For ->qlen_notify(), it is special because it is called on ->enqueue()
path, but we already hold qdisc tree lock there, and we hold this
tree lock when graft or delete the class too, so it should not be gone
or changed until we release the tree lock.
Therefore, this patch removes ->get() and ->put(), but:
1) Adds a new ->find() to find the pointer to a class by classid, no
refcnt.
2) Move the original class destroy upon the last refcnt into ->delete(),
right after releasing tree lock. This is fine because the class is
already removed from hash when holding the lock.
For those who also use ->put() as ->unbind(), just rename them to reflect
this change.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-25 07:51:29 +08:00
|
|
|
qfq_destroy_class(sch, cl);
|
|
|
|
return 0;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
net_sched: remove tc class reference counting
For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:
1) For class modification and dumping paths, we already hold RTNL lock,
so all of these ->get(),->change(),->put() are atomic.
2) For filter bindiing/unbinding, we use other reference counter than
this one, and they should have RTNL lock too.
3) For ->qlen_notify(), it is special because it is called on ->enqueue()
path, but we already hold qdisc tree lock there, and we hold this
tree lock when graft or delete the class too, so it should not be gone
or changed until we release the tree lock.
Therefore, this patch removes ->get() and ->put(), but:
1) Adds a new ->find() to find the pointer to a class by classid, no
refcnt.
2) Move the original class destroy upon the last refcnt into ->delete(),
right after releasing tree lock. This is fine because the class is
already removed from hash when holding the lock.
For those who also use ->put() as ->unbind(), just rename them to reflect
this change.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-25 07:51:29 +08:00
|
|
|
static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
net_sched: remove tc class reference counting
For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:
1) For class modification and dumping paths, we already hold RTNL lock,
so all of these ->get(),->change(),->put() are atomic.
2) For filter bindiing/unbinding, we use other reference counter than
this one, and they should have RTNL lock too.
3) For ->qlen_notify(), it is special because it is called on ->enqueue()
path, but we already hold qdisc tree lock there, and we hold this
tree lock when graft or delete the class too, so it should not be gone
or changed until we release the tree lock.
Therefore, this patch removes ->get() and ->put(), but:
1) Adds a new ->find() to find the pointer to a class by classid, no
refcnt.
2) Move the original class destroy upon the last refcnt into ->delete(),
right after releasing tree lock. This is fine because the class is
already removed from hash when holding the lock.
For those who also use ->put() as ->unbind(), just rename them to reflect
this change.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-25 07:51:29 +08:00
|
|
|
return (unsigned long)qfq_find_class(sch, classid);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2017-12-21 01:35:16 +08:00
|
|
|
static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
|
|
|
|
struct netlink_ext_ack *extack)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
|
|
|
|
if (cl)
|
|
|
|
return NULL;
|
|
|
|
|
2017-05-17 17:07:55 +08:00
|
|
|
return q->block;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
|
|
|
|
u32 classid)
|
|
|
|
{
|
|
|
|
struct qfq_class *cl = qfq_find_class(sch, classid);
|
|
|
|
|
|
|
|
if (cl != NULL)
|
|
|
|
cl->filter_cnt++;
|
|
|
|
|
|
|
|
return (unsigned long)cl;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
|
|
|
|
{
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
|
|
|
|
cl->filter_cnt--;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
|
2017-12-21 01:35:17 +08:00
|
|
|
struct Qdisc *new, struct Qdisc **old,
|
|
|
|
struct netlink_ext_ack *extack)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
|
|
|
|
if (new == NULL) {
|
2017-12-21 01:35:21 +08:00
|
|
|
new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
|
|
|
|
cl->common.classid, NULL);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (new == NULL)
|
|
|
|
new = &noop_qdisc;
|
|
|
|
}
|
|
|
|
|
2016-02-26 06:55:00 +08:00
|
|
|
*old = qdisc_replace(sch, new, &cl->qdisc);
|
2011-04-04 13:30:58 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
|
|
|
|
{
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
|
|
|
|
return cl->qdisc;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
|
|
|
|
struct sk_buff *skb, struct tcmsg *tcm)
|
|
|
|
{
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
struct nlattr *nest;
|
|
|
|
|
|
|
|
tcm->tcm_parent = TC_H_ROOT;
|
|
|
|
tcm->tcm_handle = cl->common.classid;
|
|
|
|
tcm->tcm_info = cl->qdisc->handle;
|
|
|
|
|
|
|
|
nest = nla_nest_start(skb, TCA_OPTIONS);
|
|
|
|
if (nest == NULL)
|
|
|
|
goto nla_put_failure;
|
2012-11-23 19:03:19 +08:00
|
|
|
if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
|
|
|
|
nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
|
2012-03-29 17:11:39 +08:00
|
|
|
goto nla_put_failure;
|
2011-04-04 13:30:58 +08:00
|
|
|
return nla_nest_end(skb, nest);
|
|
|
|
|
|
|
|
nla_put_failure:
|
|
|
|
nla_nest_cancel(skb, nest);
|
|
|
|
return -EMSGSIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
|
|
|
|
struct gnet_dump *d)
|
|
|
|
{
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
struct tc_qfq_stats xstats;
|
|
|
|
|
|
|
|
memset(&xstats, 0, sizeof(xstats));
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
xstats.weight = cl->agg->class_weight;
|
|
|
|
xstats.lmax = cl->agg->lmax;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2016-06-07 00:37:16 +08:00
|
|
|
if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
|
|
|
|
d, NULL, &cl->bstats) < 0 ||
|
2016-12-05 01:48:16 +08:00
|
|
|
gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
|
2019-03-28 23:53:12 +08:00
|
|
|
qdisc_qstats_copy(d, cl->qdisc) < 0)
|
2011-04-04 13:30:58 +08:00
|
|
|
return -1;
|
|
|
|
|
|
|
|
return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl;
|
|
|
|
unsigned int i;
|
|
|
|
|
|
|
|
if (arg->stop)
|
|
|
|
return;
|
|
|
|
|
|
|
|
for (i = 0; i < q->clhash.hashsize; i++) {
|
hlist: drop the node parameter from iterators
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 09:06:00 +08:00
|
|
|
hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
|
2011-04-04 13:30:58 +08:00
|
|
|
if (arg->count < arg->skip) {
|
|
|
|
arg->count++;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
|
|
|
|
arg->stop = 1;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
arg->count++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
|
|
|
|
int *qerr)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl;
|
|
|
|
struct tcf_result res;
|
2014-09-13 11:05:27 +08:00
|
|
|
struct tcf_proto *fl;
|
2011-04-04 13:30:58 +08:00
|
|
|
int result;
|
|
|
|
|
|
|
|
if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
|
|
|
|
pr_debug("qfq_classify: found %d\n", skb->priority);
|
|
|
|
cl = qfq_find_class(sch, skb->priority);
|
|
|
|
if (cl != NULL)
|
|
|
|
return cl;
|
|
|
|
}
|
|
|
|
|
|
|
|
*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
|
2014-09-13 11:05:27 +08:00
|
|
|
fl = rcu_dereference_bh(q->filter_list);
|
2017-05-17 17:07:54 +08:00
|
|
|
result = tcf_classify(skb, fl, &res, false);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (result >= 0) {
|
|
|
|
#ifdef CONFIG_NET_CLS_ACT
|
|
|
|
switch (result) {
|
|
|
|
case TC_ACT_QUEUED:
|
|
|
|
case TC_ACT_STOLEN:
|
2017-06-06 20:12:02 +08:00
|
|
|
case TC_ACT_TRAP:
|
2011-04-04 13:30:58 +08:00
|
|
|
*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
|
2017-10-20 05:28:24 +08:00
|
|
|
/* fall through */
|
2011-04-04 13:30:58 +08:00
|
|
|
case TC_ACT_SHOT:
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
cl = (struct qfq_class *)res.class;
|
|
|
|
if (cl == NULL)
|
|
|
|
cl = qfq_find_class(sch, res.classid);
|
|
|
|
return cl;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Generic comparison function, handling wraparound. */
|
|
|
|
static inline int qfq_gt(u64 a, u64 b)
|
|
|
|
{
|
|
|
|
return (s64)(a - b) > 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Round a precise timestamp to its slotted value. */
|
|
|
|
static inline u64 qfq_round_down(u64 ts, unsigned int shift)
|
|
|
|
{
|
|
|
|
return ts & ~((1ULL << shift) - 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* return the pointer to the group with lowest index in the bitmap */
|
|
|
|
static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
|
|
|
|
unsigned long bitmap)
|
|
|
|
{
|
|
|
|
int index = __ffs(bitmap);
|
|
|
|
return &q->groups[index];
|
|
|
|
}
|
|
|
|
/* Calculate a mask to mimic what would be ffs_from(). */
|
|
|
|
static inline unsigned long mask_from(unsigned long bitmap, int from)
|
|
|
|
{
|
|
|
|
return bitmap & ~((1UL << from) - 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The state computation relies on ER=0, IR=1, EB=2, IB=3
|
|
|
|
* First compute eligibility comparing grp->S, q->V,
|
|
|
|
* then check if someone is blocking us and possibly add EB
|
|
|
|
*/
|
|
|
|
static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
|
|
|
|
{
|
|
|
|
/* if S > V we are not eligible */
|
|
|
|
unsigned int state = qfq_gt(grp->S, q->V);
|
|
|
|
unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
|
|
|
|
struct qfq_group *next;
|
|
|
|
|
|
|
|
if (mask) {
|
|
|
|
next = qfq_ffs(q, mask);
|
|
|
|
if (qfq_gt(grp->F, next->F))
|
|
|
|
state |= EB;
|
|
|
|
}
|
|
|
|
|
|
|
|
return state;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* In principle
|
|
|
|
* q->bitmaps[dst] |= q->bitmaps[src] & mask;
|
|
|
|
* q->bitmaps[src] &= ~mask;
|
|
|
|
* but we should make sure that src != dst
|
|
|
|
*/
|
|
|
|
static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
|
|
|
|
int src, int dst)
|
|
|
|
{
|
|
|
|
q->bitmaps[dst] |= q->bitmaps[src] & mask;
|
|
|
|
q->bitmaps[src] &= ~mask;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
|
|
|
|
{
|
|
|
|
unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
|
|
|
|
struct qfq_group *next;
|
|
|
|
|
|
|
|
if (mask) {
|
|
|
|
next = qfq_ffs(q, mask);
|
|
|
|
if (!qfq_gt(next->F, old_F))
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
mask = (1UL << index) - 1;
|
|
|
|
qfq_move_groups(q, mask, EB, ER);
|
|
|
|
qfq_move_groups(q, mask, IB, IR);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* perhaps
|
|
|
|
*
|
|
|
|
old_V ^= q->V;
|
2012-11-23 19:03:19 +08:00
|
|
|
old_V >>= q->min_slot_shift;
|
2011-04-04 13:30:58 +08:00
|
|
|
if (old_V) {
|
|
|
|
...
|
|
|
|
}
|
|
|
|
*
|
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_make_eligible(struct qfq_sched *q)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
unsigned long vslot = q->V >> q->min_slot_shift;
|
|
|
|
unsigned long old_vslot = q->oldV >> q->min_slot_shift;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
if (vslot != old_vslot) {
|
2013-07-10 21:46:08 +08:00
|
|
|
unsigned long mask;
|
|
|
|
int last_flip_pos = fls(vslot ^ old_vslot);
|
|
|
|
|
|
|
|
if (last_flip_pos > 31) /* higher than the number of groups */
|
|
|
|
mask = ~0UL; /* make all groups eligible */
|
|
|
|
else
|
|
|
|
mask = (1UL << last_flip_pos) - 1;
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
qfq_move_groups(q, mask, IR, ER);
|
|
|
|
qfq_move_groups(q, mask, IB, EB);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
* The index of the slot in which the input aggregate agg is to be
|
|
|
|
* inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
|
|
|
|
* and not a '-1' because the start time of the group may be moved
|
|
|
|
* backward by one slot after the aggregate has been inserted, and
|
|
|
|
* this would cause non-empty slots to be right-shifted by one
|
|
|
|
* position.
|
|
|
|
*
|
|
|
|
* QFQ+ fully satisfies this bound to the slot index if the parameters
|
|
|
|
* of the classes are not changed dynamically, and if QFQ+ never
|
|
|
|
* happens to postpone the service of agg unjustly, i.e., it never
|
|
|
|
* happens that the aggregate becomes backlogged and eligible, or just
|
|
|
|
* eligible, while an aggregate with a higher approximated finish time
|
|
|
|
* is being served. In particular, in this case QFQ+ guarantees that
|
|
|
|
* the timestamps of agg are low enough that the slot index is never
|
|
|
|
* higher than 2. Unfortunately, QFQ+ cannot provide the same
|
|
|
|
* guarantee if it happens to unjustly postpone the service of agg, or
|
|
|
|
* if the parameters of some class are changed.
|
|
|
|
*
|
|
|
|
* As for the first event, i.e., an out-of-order service, the
|
|
|
|
* upper bound to the slot index guaranteed by QFQ+ grows to
|
|
|
|
* 2 +
|
|
|
|
* QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
|
|
|
|
* (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
*
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
* The following function deals with this problem by backward-shifting
|
|
|
|
* the timestamps of agg, if needed, so as to guarantee that the slot
|
|
|
|
* index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
|
|
|
|
* cause the service of other aggregates to be postponed, yet the
|
|
|
|
* worst-case guarantees of these aggregates are not violated. In
|
|
|
|
* fact, in case of no out-of-order service, the timestamps of agg
|
|
|
|
* would have been even lower than they are after the backward shift,
|
|
|
|
* because QFQ+ would have guaranteed a maximum value equal to 2 for
|
|
|
|
* the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
|
|
|
|
* service is postponed because of the backward-shift would have
|
|
|
|
* however waited for the service of agg before being served.
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
*
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
* The other event that may cause the slot index to be higher than 2
|
|
|
|
* for agg is a recent change of the parameters of some class. If the
|
|
|
|
* weight of a class is increased or the lmax (max_pkt_size) of the
|
|
|
|
* class is decreased, then a new aggregate with smaller slot size
|
|
|
|
* than the original parent aggregate of the class may happen to be
|
|
|
|
* activated. The activation of this aggregate should be properly
|
|
|
|
* delayed to when the service of the class has finished in the ideal
|
|
|
|
* system tracked by QFQ+. If the activation of the aggregate is not
|
|
|
|
* delayed to this reference time instant, then this aggregate may be
|
|
|
|
* unjustly served before other aggregates waiting for service. This
|
|
|
|
* may cause the above bound to the slot index to be violated for some
|
|
|
|
* of these unlucky aggregates.
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Instead of delaying the activation of the new aggregate, which is
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
* quite complex, the above-discussed capping of the slot index is
|
|
|
|
* used to handle also the consequences of a change of the parameters
|
|
|
|
* of a class.
|
2011-04-04 13:30:58 +08:00
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
|
2011-04-04 13:30:58 +08:00
|
|
|
u64 roundedS)
|
|
|
|
{
|
|
|
|
u64 slot = (roundedS - grp->S) >> grp->slot_shift;
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
unsigned int i; /* slot index in the bucket list */
|
|
|
|
|
|
|
|
if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
|
|
|
|
u64 deltaS = roundedS - grp->S -
|
|
|
|
((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
|
2012-11-23 19:03:19 +08:00
|
|
|
agg->S -= deltaS;
|
|
|
|
agg->F -= deltaS;
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
slot = QFQ_MAX_SLOTS - 2;
|
|
|
|
}
|
|
|
|
|
|
|
|
i = (grp->front + slot) % QFQ_MAX_SLOTS;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
hlist_add_head(&agg->next, &grp->slots[i]);
|
2011-04-04 13:30:58 +08:00
|
|
|
__set_bit(slot, &grp->full_slots);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Maybe introduce hlist_first_entry?? */
|
2012-11-23 19:03:19 +08:00
|
|
|
static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
return hlist_entry(grp->slots[grp->front].first,
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate, next);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* remove the entry from the slot
|
|
|
|
*/
|
|
|
|
static void qfq_front_slot_remove(struct qfq_group *grp)
|
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *agg = qfq_slot_head(grp);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
BUG_ON(!agg);
|
|
|
|
hlist_del(&agg->next);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (hlist_empty(&grp->slots[grp->front]))
|
|
|
|
__clear_bit(0, &grp->full_slots);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Returns the first aggregate in the first non-empty bucket of the
|
|
|
|
* group. As a side effect, adjusts the bucket list so the first
|
|
|
|
* non-empty bucket is at position 0 in full_slots.
|
2011-04-04 13:30:58 +08:00
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
unsigned int i;
|
|
|
|
|
|
|
|
pr_debug("qfq slot_scan: grp %u full %#lx\n",
|
|
|
|
grp->index, grp->full_slots);
|
|
|
|
|
|
|
|
if (grp->full_slots == 0)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
i = __ffs(grp->full_slots); /* zero based */
|
|
|
|
if (i > 0) {
|
|
|
|
grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
|
|
|
|
grp->full_slots >>= i;
|
|
|
|
}
|
|
|
|
|
|
|
|
return qfq_slot_head(grp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* adjust the bucket list. When the start time of a group decreases,
|
|
|
|
* we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
|
|
|
|
* move the objects. The mask of occupied slots must be shifted
|
|
|
|
* because we use ffs() to find the first non-empty slot.
|
|
|
|
* This covers decreases in the group's start time, but what about
|
|
|
|
* increases of the start time ?
|
|
|
|
* Here too we should make sure that i is less than 32
|
|
|
|
*/
|
|
|
|
static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
|
|
|
|
{
|
|
|
|
unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
|
|
|
|
|
|
|
|
grp->full_slots <<= i;
|
|
|
|
grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
|
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_update_eligible(struct qfq_sched *q)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
struct qfq_group *grp;
|
|
|
|
unsigned long ineligible;
|
|
|
|
|
|
|
|
ineligible = q->bitmaps[IR] | q->bitmaps[IB];
|
|
|
|
if (ineligible) {
|
|
|
|
if (!q->bitmaps[ER]) {
|
|
|
|
grp = qfq_ffs(q, ineligible);
|
|
|
|
if (qfq_gt(grp->S, q->V))
|
|
|
|
q->V = grp->S;
|
|
|
|
}
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_make_eligible(q);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Dequeue head packet of the head class in the DRR queue of the aggregate. */
|
|
|
|
static void agg_dequeue(struct qfq_aggregate *agg,
|
|
|
|
struct qfq_class *cl, unsigned int len)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
qdisc_dequeue_peeked(cl->qdisc);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
cl->deficit -= (int) len;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
|
|
|
|
list_del(&cl->alist);
|
|
|
|
else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
|
|
|
|
cl->deficit += agg->lmax;
|
|
|
|
list_move_tail(&cl->alist, &agg->active);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
2012-11-23 19:03:19 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
|
|
|
|
struct qfq_class **cl,
|
|
|
|
unsigned int *len)
|
|
|
|
{
|
|
|
|
struct sk_buff *skb;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
*cl = list_first_entry(&agg->active, struct qfq_class, alist);
|
|
|
|
skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
|
|
|
|
if (skb == NULL)
|
|
|
|
WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
|
|
|
|
else
|
|
|
|
*len = qdisc_pkt_len(skb);
|
|
|
|
|
|
|
|
return skb;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Update F according to the actual service received by the aggregate. */
|
|
|
|
static inline void charge_actual_service(struct qfq_aggregate *agg)
|
|
|
|
{
|
2013-03-05 16:04:57 +08:00
|
|
|
/* Compute the service received by the aggregate, taking into
|
|
|
|
* account that, after decreasing the number of classes in
|
|
|
|
* agg, it may happen that
|
|
|
|
* agg->initial_budget - agg->budget > agg->bugdetmax
|
|
|
|
*/
|
|
|
|
u32 service_received = min(agg->budgetmax,
|
|
|
|
agg->initial_budget - agg->budget);
|
2012-11-23 19:03:19 +08:00
|
|
|
|
|
|
|
agg->F = agg->S + (u64)service_received * agg->inv_w;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2013-07-10 21:46:09 +08:00
|
|
|
/* Assign a reasonable start time for a new aggregate in group i.
|
|
|
|
* Admissible values for \hat(F) are multiples of \sigma_i
|
|
|
|
* no greater than V+\sigma_i . Larger values mean that
|
|
|
|
* we had a wraparound so we consider the timestamp to be stale.
|
|
|
|
*
|
|
|
|
* If F is not stale and F >= V then we set S = F.
|
|
|
|
* Otherwise we should assign S = V, but this may violate
|
|
|
|
* the ordering in EB (see [2]). So, if we have groups in ER,
|
|
|
|
* set S to the F_j of the first group j which would be blocking us.
|
|
|
|
* We are guaranteed not to move S backward because
|
|
|
|
* otherwise our group i would still be blocked.
|
|
|
|
*/
|
|
|
|
static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
|
|
|
|
{
|
|
|
|
unsigned long mask;
|
|
|
|
u64 limit, roundedF;
|
|
|
|
int slot_shift = agg->grp->slot_shift;
|
|
|
|
|
|
|
|
roundedF = qfq_round_down(agg->F, slot_shift);
|
|
|
|
limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
|
|
|
|
|
|
|
|
if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
|
|
|
|
/* timestamp was stale */
|
|
|
|
mask = mask_from(q->bitmaps[ER], agg->grp->index);
|
|
|
|
if (mask) {
|
|
|
|
struct qfq_group *next = qfq_ffs(q, mask);
|
|
|
|
if (qfq_gt(roundedF, next->F)) {
|
|
|
|
if (qfq_gt(limit, next->F))
|
|
|
|
agg->S = next->F;
|
|
|
|
else /* preserve timestamp correctness */
|
|
|
|
agg->S = limit;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
agg->S = q->V;
|
|
|
|
} else /* timestamp is not stale */
|
|
|
|
agg->S = agg->F;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Update the timestamps of agg before scheduling/rescheduling it for
|
|
|
|
* service. In particular, assign to agg->F its maximum possible
|
|
|
|
* value, i.e., the virtual finish time with which the aggregate
|
|
|
|
* should be labeled if it used all its budget once in service.
|
|
|
|
*/
|
|
|
|
static inline void
|
|
|
|
qfq_update_agg_ts(struct qfq_sched *q,
|
|
|
|
struct qfq_aggregate *agg, enum update_reason reason)
|
|
|
|
{
|
|
|
|
if (reason != requeue)
|
|
|
|
qfq_update_start(q, agg);
|
|
|
|
else /* just charge agg for the service received */
|
|
|
|
agg->S = agg->F;
|
|
|
|
|
|
|
|
agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
|
|
|
|
}
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
|
|
|
|
static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
|
2011-04-04 13:30:58 +08:00
|
|
|
struct qfq_class *cl;
|
2012-11-23 19:03:19 +08:00
|
|
|
struct sk_buff *skb = NULL;
|
|
|
|
/* next-packet len, 0 means no more active classes in in-service agg */
|
|
|
|
unsigned int len = 0;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (in_serv_agg == NULL)
|
2011-04-04 13:30:58 +08:00
|
|
|
return NULL;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (!list_empty(&in_serv_agg->active))
|
|
|
|
skb = qfq_peek_skb(in_serv_agg, &cl, &len);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/*
|
|
|
|
* If there are no active classes in the in-service aggregate,
|
|
|
|
* or if the aggregate has not enough budget to serve its next
|
|
|
|
* class, then choose the next aggregate to serve.
|
|
|
|
*/
|
|
|
|
if (len == 0 || in_serv_agg->budget < len) {
|
|
|
|
charge_actual_service(in_serv_agg);
|
|
|
|
|
|
|
|
/* recharge the budget of the aggregate */
|
|
|
|
in_serv_agg->initial_budget = in_serv_agg->budget =
|
|
|
|
in_serv_agg->budgetmax;
|
|
|
|
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
if (!list_empty(&in_serv_agg->active)) {
|
2012-11-23 19:03:19 +08:00
|
|
|
/*
|
|
|
|
* Still active: reschedule for
|
|
|
|
* service. Possible optimization: if no other
|
|
|
|
* aggregate is active, then there is no point
|
|
|
|
* in rescheduling this aggregate, and we can
|
|
|
|
* just keep it as the in-service one. This
|
|
|
|
* should be however a corner case, and to
|
|
|
|
* handle it, we would need to maintain an
|
|
|
|
* extra num_active_aggs field.
|
|
|
|
*/
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
qfq_update_agg_ts(q, in_serv_agg, requeue);
|
|
|
|
qfq_schedule_agg(q, in_serv_agg);
|
|
|
|
} else if (sch->q.qlen == 0) { /* no aggregate to serve */
|
2012-11-23 19:03:19 +08:00
|
|
|
q->in_serv_agg = NULL;
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we get here, there are other aggregates queued:
|
|
|
|
* choose the new aggregate to serve.
|
|
|
|
*/
|
|
|
|
in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
|
|
|
|
skb = qfq_peek_skb(in_serv_agg, &cl, &len);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
2012-11-23 19:03:19 +08:00
|
|
|
if (!skb)
|
|
|
|
return NULL;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2016-09-19 07:22:47 +08:00
|
|
|
qdisc_qstats_backlog_dec(sch, skb);
|
2011-04-04 13:30:58 +08:00
|
|
|
sch->q.qlen--;
|
|
|
|
qdisc_bstats_update(sch, skb);
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
agg_dequeue(in_serv_agg, cl, len);
|
pkt_sched: sch_qfq: prevent budget from wrapping around after a dequeue
Aggregate budgets are computed so as to guarantee that, after an
aggregate has been selected for service, that aggregate has enough
budget to serve at least one maximum-size packet for the classes it
contains. For this reason, after a new aggregate has been selected
for service, its next packet is immediately dequeued, without any
further control.
The maximum packet size for a class, lmax, can be changed through
qfq_change_class(). In case the user sets lmax to a lower value than
the the size of some of the still-to-arrive packets, QFQ+ will
automatically push up lmax as it enqueues these packets. This
automatic push up is likely to happen with TSO/GSO.
In any case, if lmax is assigned a lower value than the size of some
of the packets already enqueued for the class, then the following
problem may occur: the size of the next packet to dequeue for the
class may happen to be larger than lmax, after the aggregate to which
the class belongs has been just selected for service. In this case,
even the budget of the aggregate, which is an unsigned value, may be
lower than the size of the next packet to dequeue. After dequeueing
this packet and subtracting its size from the budget, the latter would
wrap around.
This fix prevents the budget from wrapping around after any packet
dequeue.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:05:00 +08:00
|
|
|
/* If lmax is lowered, through qfq_change_class, for a class
|
|
|
|
* owning pending packets with larger size than the new value
|
|
|
|
* of lmax, then the following condition may hold.
|
|
|
|
*/
|
|
|
|
if (unlikely(in_serv_agg->budget < len))
|
|
|
|
in_serv_agg->budget = 0;
|
|
|
|
else
|
|
|
|
in_serv_agg->budget -= len;
|
|
|
|
|
pkt_sched: sch_qfq: remove a source of high packet delay/jitter
QFQ+ inherits from QFQ a design choice that may cause a high packet
delay/jitter and a severe short-term unfairness. As QFQ, QFQ+ uses a
special quantity, the system virtual time, to track the service
provided by the ideal system it approximates. When a packet is
dequeued, this quantity must be incremented by the size of the packet,
divided by the sum of the weights of the aggregates waiting to be
served. Tracking this sum correctly is a non-trivial task, because, to
preserve tight service guarantees, the decrement of this sum must be
delayed in a special way [1]: this sum can be decremented only after
that its value would decrease also in the ideal system approximated by
QFQ+. For efficiency, QFQ+ keeps track only of the 'instantaneous'
weight sum, increased and decreased immediately as the weight of an
aggregate changes, and as an aggregate is created or destroyed (which,
in its turn, happens as a consequence of some class being
created/destroyed/changed). However, to avoid the problems caused to
service guarantees by these immediate decreases, QFQ+ increments the
system virtual time using the maximum value allowed for the weight
sum, 2^10, in place of the dynamic, instantaneous value. The
instantaneous value of the weight sum is used only to check whether a
request of weight increase or a class creation can be satisfied.
Unfortunately, the problems caused by this choice are worse than the
temporary degradation of the service guarantees that may occur, when a
class is changed or destroyed, if the instantaneous value of the
weight sum was used to update the system virtual time. In fact, the
fraction of the link bandwidth guaranteed by QFQ+ to each aggregate is
equal to the ratio between the weight of the aggregate and the sum of
the weights of the competing aggregates. The packet delay guaranteed
to the aggregate is instead inversely proportional to the guaranteed
bandwidth. By using the maximum possible value, and not the actual
value of the weight sum, QFQ+ provides each aggregate with the worst
possible service guarantees, and not with service guarantees related
to the actual set of competing aggregates. To see the consequences of
this fact, consider the following simple example.
Suppose that only the following aggregates are backlogged, i.e., that
only the classes in the following aggregates have packets to transmit:
one aggregate with weight 10, say A, and ten aggregates with weight 1,
say B1, B2, ..., B10. In particular, suppose that these aggregates are
always backlogged. Given the weight distribution, the smoothest and
fairest service order would be:
A B1 A B2 A B3 A B4 A B5 A B6 A B7 A B8 A B9 A B10 A B1 A B2 ...
QFQ+ would provide exactly this optimal service if it used the actual
value for the weight sum instead of the maximum possible value, i.e.,
11 instead of 2^10. In contrast, since QFQ+ uses the latter value, it
serves aggregates as follows (easy to prove and to reproduce
experimentally):
A B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 A A A A A A A A A A B1 B2 ... B10 A A ...
By replacing 10 with N in the above example, and by increasing N, one
can increase at will the maximum packet delay and the jitter
experienced by the classes in aggregate A.
This patch addresses this issue by just using the above
'instantaneous' value of the weight sum, instead of the maximum
possible value, when updating the system virtual time. After the
instantaneous weight sum is decreased, QFQ+ may deviate from the ideal
service for a time interval in the order of the time to serve one
maximum-size packet for each backlogged class. The worst-case extent
of the deviation exhibited by QFQ+ during this time interval [1] is
basically the same as of the deviation described above (but, without
this patch, QFQ+ suffers from such a deviation all the time). Finally,
this patch modifies the comment to the function qfq_slot_insert, to
make it coherent with the fact that the weight sum used by QFQ+ can
now be lower than the maximum possible value.
[1] P. Valente, "Extending WF2Q+ to support a dynamic traffic mix",
Proceedings of AAA-IDEA'05, June 2005.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-07-16 14:52:30 +08:00
|
|
|
q->V += (u64)len * q->iwsum;
|
2011-04-04 13:30:58 +08:00
|
|
|
pr_debug("qfq dequeue: len %u F %lld now %lld\n",
|
2012-11-23 19:03:19 +08:00
|
|
|
len, (unsigned long long) in_serv_agg->F,
|
|
|
|
(unsigned long long) q->V);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
return skb;
|
|
|
|
}
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
|
|
|
|
{
|
|
|
|
struct qfq_group *grp;
|
|
|
|
struct qfq_aggregate *agg, *new_front_agg;
|
|
|
|
u64 old_F;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_update_eligible(q);
|
|
|
|
q->oldV = q->V;
|
|
|
|
|
|
|
|
if (!q->bitmaps[ER])
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
grp = qfq_ffs(q, q->bitmaps[ER]);
|
|
|
|
old_F = grp->F;
|
|
|
|
|
|
|
|
agg = qfq_slot_head(grp);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* agg starts to be served, remove it from schedule */
|
|
|
|
qfq_front_slot_remove(grp);
|
|
|
|
|
|
|
|
new_front_agg = qfq_slot_scan(grp);
|
|
|
|
|
|
|
|
if (new_front_agg == NULL) /* group is now inactive, remove from ER */
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[ER]);
|
|
|
|
else {
|
|
|
|
u64 roundedS = qfq_round_down(new_front_agg->S,
|
|
|
|
grp->slot_shift);
|
|
|
|
unsigned int s;
|
|
|
|
|
|
|
|
if (grp->S == roundedS)
|
|
|
|
return agg;
|
|
|
|
grp->S = roundedS;
|
|
|
|
grp->F = roundedS + (2ULL << grp->slot_shift);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[ER]);
|
|
|
|
s = qfq_calc_state(q, grp);
|
|
|
|
__set_bit(grp->index, &q->bitmaps[s]);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_unblock_groups(q, grp->index, old_F);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
return agg;
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2016-06-22 14:16:49 +08:00
|
|
|
static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
|
|
|
|
struct sk_buff **to_free)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
2019-01-10 00:09:42 +08:00
|
|
|
unsigned int len = qdisc_pkt_len(skb), gso_segs;
|
2011-04-04 13:30:58 +08:00
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl;
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *agg;
|
2012-09-28 06:35:47 +08:00
|
|
|
int err = 0;
|
2019-01-10 00:09:43 +08:00
|
|
|
bool first;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
cl = qfq_classify(skb, sch, &err);
|
|
|
|
if (cl == NULL) {
|
|
|
|
if (err & __NET_XMIT_BYPASS)
|
2014-09-29 02:53:29 +08:00
|
|
|
qdisc_qstats_drop(sch);
|
2017-09-04 14:21:12 +08:00
|
|
|
__qdisc_drop(skb, to_free);
|
2011-04-04 13:30:58 +08:00
|
|
|
return err;
|
|
|
|
}
|
|
|
|
pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
|
|
|
|
|
2019-01-10 00:09:42 +08:00
|
|
|
if (unlikely(cl->agg->lmax < len)) {
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
|
2019-01-10 00:09:42 +08:00
|
|
|
cl->agg->lmax, len, cl->common.classid);
|
|
|
|
err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
|
2016-06-09 05:23:01 +08:00
|
|
|
if (err) {
|
|
|
|
cl->qstats.drops++;
|
2016-06-22 14:16:49 +08:00
|
|
|
return qdisc_drop(skb, sch, to_free);
|
2016-06-09 05:23:01 +08:00
|
|
|
}
|
pkt_sched: enable QFQ to support TSO/GSO
If the max packet size for some class (configured through tc) is
violated by the actual size of the packets of that class, then QFQ
would not schedule classes correctly, and the data structures
implementing the bucket lists may get corrupted. This problem occurs
with TSO/GSO even if the max packet size is set to the MTU, and is,
e.g., the cause of the failure reported in [1]. Two patches have been
proposed to solve this problem in [2], one of them is a preliminary
version of this patch.
This patch addresses the above issues by: 1) setting QFQ parameters to
proper values for supporting TSO/GSO (in particular, setting the
maximum possible packet size to 64KB), 2) automatically increasing the
max packet size for a class, lmax, when a packet with a larger size
than the current value of lmax arrives.
The drawback of the first point is that the maximum weight for a class
is now limited to 4096, which is equal to 1/16 of the maximum weight
sum.
Finally, this patch also forcibly caps the timestamps of a class if
they are too high to be stored in the bucket list. This capping, taken
from QFQ+ [3], handles the unfrequent case described in the comment to
the function slot_insert.
[1] http://marc.info/?l=linux-netdev&m=134968777902077&w=2
[2] http://marc.info/?l=linux-netdev&m=135096573507936&w=2
[3] http://marc.info/?l=linux-netdev&m=134902691421670&w=2
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Tested-by: Cong Wang <amwang@redhat.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Acked-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2012-11-06 04:29:24 +08:00
|
|
|
}
|
|
|
|
|
2019-01-10 00:09:42 +08:00
|
|
|
gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
|
2019-01-10 00:09:43 +08:00
|
|
|
first = !cl->qdisc->q.qlen;
|
2016-06-22 14:16:49 +08:00
|
|
|
err = qdisc_enqueue(skb, cl->qdisc, to_free);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (unlikely(err != NET_XMIT_SUCCESS)) {
|
|
|
|
pr_debug("qfq_enqueue: enqueue failed %d\n", err);
|
|
|
|
if (net_xmit_drop_count(err)) {
|
|
|
|
cl->qstats.drops++;
|
2014-09-29 02:53:29 +08:00
|
|
|
qdisc_qstats_drop(sch);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
2019-01-10 00:09:42 +08:00
|
|
|
cl->bstats.bytes += len;
|
|
|
|
cl->bstats.packets += gso_segs;
|
|
|
|
sch->qstats.backlog += len;
|
2011-04-04 13:30:58 +08:00
|
|
|
++sch->q.qlen;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
agg = cl->agg;
|
|
|
|
/* if the queue was not empty, then done here */
|
2019-01-10 00:09:43 +08:00
|
|
|
if (!first) {
|
2012-11-23 19:03:19 +08:00
|
|
|
if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
|
|
|
|
list_first_entry(&agg->active, struct qfq_class, alist)
|
2019-01-10 00:09:42 +08:00
|
|
|
== cl && cl->deficit < len)
|
2012-11-23 19:03:19 +08:00
|
|
|
list_move_tail(&cl->alist, &agg->active);
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
return err;
|
2012-11-23 19:03:19 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* schedule class for service within the aggregate */
|
|
|
|
cl->deficit = agg->lmax;
|
|
|
|
list_add_tail(&cl->alist, &agg->active);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
|
|
|
|
q->in_serv_agg == agg)
|
|
|
|
return err; /* non-empty or in service, nothing else to do */
|
2012-11-23 19:03:19 +08:00
|
|
|
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
qfq_activate_agg(q, agg, enqueue);
|
2012-08-07 15:27:25 +08:00
|
|
|
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Schedule aggregate according to its timestamps.
|
2012-08-07 15:27:25 +08:00
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
|
2012-08-07 15:27:25 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_group *grp = agg->grp;
|
2012-08-07 15:27:25 +08:00
|
|
|
u64 roundedS;
|
|
|
|
int s;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
roundedS = qfq_round_down(agg->S, grp->slot_shift);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Insert agg in the correct bucket.
|
|
|
|
* If agg->S >= grp->S we don't need to adjust the
|
2011-04-04 13:30:58 +08:00
|
|
|
* bucket list and simply go to the insertion phase.
|
|
|
|
* Otherwise grp->S is decreasing, we must make room
|
|
|
|
* in the bucket list, and also recompute the group state.
|
|
|
|
* Finally, if there were no flows in this group and nobody
|
|
|
|
* was in ER make sure to adjust V.
|
|
|
|
*/
|
|
|
|
if (grp->full_slots) {
|
2012-11-23 19:03:19 +08:00
|
|
|
if (!qfq_gt(grp->S, agg->S))
|
2011-04-04 13:30:58 +08:00
|
|
|
goto skip_update;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* create a slot for this agg->S */
|
2011-04-04 13:30:58 +08:00
|
|
|
qfq_slot_rotate(grp, roundedS);
|
|
|
|
/* group was surely ineligible, remove */
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IR]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IB]);
|
2013-03-05 16:05:01 +08:00
|
|
|
} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
|
|
|
|
q->in_serv_agg == NULL)
|
2011-04-04 13:30:58 +08:00
|
|
|
q->V = roundedS;
|
|
|
|
|
|
|
|
grp->S = roundedS;
|
|
|
|
grp->F = roundedS + (2ULL << grp->slot_shift);
|
|
|
|
s = qfq_calc_state(q, grp);
|
|
|
|
__set_bit(grp->index, &q->bitmaps[s]);
|
|
|
|
|
|
|
|
pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
|
|
|
|
s, q->bitmaps[s],
|
2012-11-23 19:03:19 +08:00
|
|
|
(unsigned long long) agg->S,
|
|
|
|
(unsigned long long) agg->F,
|
2011-04-04 13:30:58 +08:00
|
|
|
(unsigned long long) q->V);
|
|
|
|
|
|
|
|
skip_update:
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_slot_insert(grp, agg, roundedS);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
/* Update agg ts and schedule agg for service */
|
|
|
|
static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
|
|
|
|
enum update_reason reason)
|
|
|
|
{
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
qfq_update_agg_ts(q, agg, reason);
|
pkt_sched: sch_qfq: serve activated aggregates immediately if the scheduler is empty
If no aggregate is in service, then the function qfq_dequeue() does
not dequeue any packet. For this reason, to guarantee QFQ+ to be work
conserving, a just-activated aggregate must be set as in service
immediately if it happens to be the only active aggregate.
This is done by the function qfq_enqueue().
Unfortunately, the function qfq_add_to_agg(), used to add a class to
an aggregate, does not perform this important additional operation.
In particular, if: 1) qfq_add_to_agg() is invoked to complete the move
of a class from a source aggregate, becoming, for this move, inactive,
to a destination aggregate, becoming instead active, and 2) the
destination aggregate becomes the only active aggregate, then this
aggregate is not however set as in service. QFQ+ remains then in a
non-work-conserving state until a new invocation of qfq_enqueue()
recovers the situation.
This fix solves the problem by moving the logic for setting an
aggregate as in service directly into the function qfq_activate_agg().
Hence, from whatever point qfq_activate_aggregate() is invoked, QFQ+
remains work conserving. Since the more-complex logic of this new
version of activate_aggregate() is not necessary, in qfq_dequeue(), to
reschedule an aggregate that finishes its budget, then the aggregate
is now rescheduled by invoking directly the functions needed.
Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
Reviewed-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-03-05 16:04:59 +08:00
|
|
|
if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
|
|
|
|
q->in_serv_agg = agg; /* start serving this aggregate */
|
|
|
|
/* update V: to be in service, agg must be eligible */
|
|
|
|
q->oldV = q->V = agg->S;
|
|
|
|
} else if (agg != q->in_serv_agg)
|
|
|
|
qfq_schedule_agg(q, agg);
|
2012-11-23 19:03:19 +08:00
|
|
|
}
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_aggregate *agg)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
unsigned int i, offset;
|
|
|
|
u64 roundedS;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
roundedS = qfq_round_down(agg->S, grp->slot_shift);
|
2011-04-04 13:30:58 +08:00
|
|
|
offset = (roundedS - grp->S) >> grp->slot_shift;
|
2012-11-23 19:03:19 +08:00
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
i = (grp->front + offset) % QFQ_MAX_SLOTS;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
hlist_del(&agg->next);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (hlist_empty(&grp->slots[i]))
|
|
|
|
__clear_bit(offset, &grp->full_slots);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2012-11-23 19:03:19 +08:00
|
|
|
* Called to forcibly deschedule an aggregate. If the aggregate is
|
|
|
|
* not in the front bucket, or if the latter has other aggregates in
|
|
|
|
* the front bucket, we can simply remove the aggregate with no other
|
|
|
|
* side effects.
|
2011-04-04 13:30:58 +08:00
|
|
|
* Otherwise we must propagate the event up.
|
|
|
|
*/
|
2012-11-23 19:03:19 +08:00
|
|
|
static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
2012-11-23 19:03:19 +08:00
|
|
|
struct qfq_group *grp = agg->grp;
|
2011-04-04 13:30:58 +08:00
|
|
|
unsigned long mask;
|
|
|
|
u64 roundedS;
|
|
|
|
int s;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (agg == q->in_serv_agg) {
|
|
|
|
charge_actual_service(agg);
|
|
|
|
q->in_serv_agg = qfq_choose_next_agg(q);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
agg->F = agg->S;
|
|
|
|
qfq_slot_remove(q, grp, agg);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
if (!grp->full_slots) {
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IR]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[EB]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IB]);
|
|
|
|
|
|
|
|
if (test_bit(grp->index, &q->bitmaps[ER]) &&
|
|
|
|
!(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
|
|
|
|
mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
|
|
|
|
if (mask)
|
|
|
|
mask = ~((1UL << __fls(mask)) - 1);
|
|
|
|
else
|
|
|
|
mask = ~0UL;
|
|
|
|
qfq_move_groups(q, mask, EB, ER);
|
|
|
|
qfq_move_groups(q, mask, IB, IR);
|
|
|
|
}
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[ER]);
|
|
|
|
} else if (hlist_empty(&grp->slots[grp->front])) {
|
2012-11-23 19:03:19 +08:00
|
|
|
agg = qfq_slot_scan(grp);
|
|
|
|
roundedS = qfq_round_down(agg->S, grp->slot_shift);
|
2011-04-04 13:30:58 +08:00
|
|
|
if (grp->S != roundedS) {
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[ER]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IR]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[EB]);
|
|
|
|
__clear_bit(grp->index, &q->bitmaps[IB]);
|
|
|
|
grp->S = roundedS;
|
|
|
|
grp->F = roundedS + (2ULL << grp->slot_shift);
|
|
|
|
s = qfq_calc_state(q, grp);
|
|
|
|
__set_bit(grp->index, &q->bitmaps[s]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl = (struct qfq_class *)arg;
|
|
|
|
|
2017-08-15 21:39:59 +08:00
|
|
|
qfq_deactivate_class(q, cl);
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
|
|
|
|
2017-12-21 01:35:13 +08:00
|
|
|
static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
|
|
|
|
struct netlink_ext_ack *extack)
|
2011-04-04 13:30:58 +08:00
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_group *grp;
|
|
|
|
int i, j, err;
|
2012-11-23 19:03:19 +08:00
|
|
|
u32 max_cl_shift, maxbudg_shift, max_classes;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2017-12-21 01:35:19 +08:00
|
|
|
err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
|
2017-05-17 17:07:55 +08:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
err = qdisc_class_hash_init(&q->clhash);
|
|
|
|
if (err < 0)
|
|
|
|
return err;
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
|
|
|
|
max_classes = QFQ_MAX_AGG_CLASSES;
|
|
|
|
else
|
|
|
|
max_classes = qdisc_dev(sch)->tx_queue_len + 1;
|
|
|
|
/* max_cl_shift = floor(log_2(max_classes)) */
|
|
|
|
max_cl_shift = __fls(max_classes);
|
|
|
|
q->max_agg_classes = 1<<max_cl_shift;
|
|
|
|
|
|
|
|
/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
|
|
|
|
maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
|
|
|
|
q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
for (i = 0; i <= QFQ_MAX_INDEX; i++) {
|
|
|
|
grp = &q->groups[i];
|
|
|
|
grp->index = i;
|
2012-11-23 19:03:19 +08:00
|
|
|
grp->slot_shift = q->min_slot_shift + i;
|
2011-04-04 13:30:58 +08:00
|
|
|
for (j = 0; j < QFQ_MAX_SLOTS; j++)
|
|
|
|
INIT_HLIST_HEAD(&grp->slots[j]);
|
|
|
|
}
|
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
INIT_HLIST_HEAD(&q->nonfull_aggs);
|
|
|
|
|
2011-04-04 13:30:58 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_reset_qdisc(struct Qdisc *sch)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl;
|
2012-11-23 19:03:19 +08:00
|
|
|
unsigned int i;
|
2011-04-04 13:30:58 +08:00
|
|
|
|
2012-11-23 19:03:19 +08:00
|
|
|
for (i = 0; i < q->clhash.hashsize; i++) {
|
hlist: drop the node parameter from iterators
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 09:06:00 +08:00
|
|
|
hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
|
2012-11-23 19:03:19 +08:00
|
|
|
if (cl->qdisc->q.qlen > 0)
|
2011-04-04 13:30:58 +08:00
|
|
|
qfq_deactivate_class(q, cl);
|
|
|
|
|
|
|
|
qdisc_reset(cl->qdisc);
|
2012-11-23 19:03:19 +08:00
|
|
|
}
|
2011-04-04 13:30:58 +08:00
|
|
|
}
|
2016-09-19 07:22:47 +08:00
|
|
|
sch->qstats.backlog = 0;
|
2011-04-04 13:30:58 +08:00
|
|
|
sch->q.qlen = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void qfq_destroy_qdisc(struct Qdisc *sch)
|
|
|
|
{
|
|
|
|
struct qfq_sched *q = qdisc_priv(sch);
|
|
|
|
struct qfq_class *cl;
|
hlist: drop the node parameter from iterators
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 09:06:00 +08:00
|
|
|
struct hlist_node *next;
|
2011-04-04 13:30:58 +08:00
|
|
|
unsigned int i;
|
|
|
|
|
2017-05-17 17:07:55 +08:00
|
|
|
tcf_block_put(q->block);
|
2011-04-04 13:30:58 +08:00
|
|
|
|
|
|
|
for (i = 0; i < q->clhash.hashsize; i++) {
|
hlist: drop the node parameter from iterators
I'm not sure why, but the hlist for each entry iterators were conceived
list_for_each_entry(pos, head, member)
The hlist ones were greedy and wanted an extra parameter:
hlist_for_each_entry(tpos, pos, head, member)
Why did they need an extra pos parameter? I'm not quite sure. Not only
they don't really need it, it also prevents the iterator from looking
exactly like the list iterator, which is unfortunate.
Besides the semantic patch, there was some manual work required:
- Fix up the actual hlist iterators in linux/list.h
- Fix up the declaration of other iterators based on the hlist ones.
- A very small amount of places were using the 'node' parameter, this
was modified to use 'obj->member' instead.
- Coccinelle didn't handle the hlist_for_each_entry_safe iterator
properly, so those had to be fixed up manually.
The semantic patch which is mostly the work of Peter Senna Tschudin is here:
@@
iterator name hlist_for_each_entry, hlist_for_each_entry_continue, hlist_for_each_entry_from, hlist_for_each_entry_rcu, hlist_for_each_entry_rcu_bh, hlist_for_each_entry_continue_rcu_bh, for_each_busy_worker, ax25_uid_for_each, ax25_for_each, inet_bind_bucket_for_each, sctp_for_each_hentry, sk_for_each, sk_for_each_rcu, sk_for_each_from, sk_for_each_safe, sk_for_each_bound, hlist_for_each_entry_safe, hlist_for_each_entry_continue_rcu, nr_neigh_for_each, nr_neigh_for_each_safe, nr_node_for_each, nr_node_for_each_safe, for_each_gfn_indirect_valid_sp, for_each_gfn_sp, for_each_host;
type T;
expression a,c,d,e;
identifier b;
statement S;
@@
-T b;
<+... when != b
(
hlist_for_each_entry(a,
- b,
c, d) S
|
hlist_for_each_entry_continue(a,
- b,
c) S
|
hlist_for_each_entry_from(a,
- b,
c) S
|
hlist_for_each_entry_rcu(a,
- b,
c, d) S
|
hlist_for_each_entry_rcu_bh(a,
- b,
c, d) S
|
hlist_for_each_entry_continue_rcu_bh(a,
- b,
c) S
|
for_each_busy_worker(a, c,
- b,
d) S
|
ax25_uid_for_each(a,
- b,
c) S
|
ax25_for_each(a,
- b,
c) S
|
inet_bind_bucket_for_each(a,
- b,
c) S
|
sctp_for_each_hentry(a,
- b,
c) S
|
sk_for_each(a,
- b,
c) S
|
sk_for_each_rcu(a,
- b,
c) S
|
sk_for_each_from
-(a, b)
+(a)
S
+ sk_for_each_from(a) S
|
sk_for_each_safe(a,
- b,
c, d) S
|
sk_for_each_bound(a,
- b,
c) S
|
hlist_for_each_entry_safe(a,
- b,
c, d, e) S
|
hlist_for_each_entry_continue_rcu(a,
- b,
c) S
|
nr_neigh_for_each(a,
- b,
c) S
|
nr_neigh_for_each_safe(a,
- b,
c, d) S
|
nr_node_for_each(a,
- b,
c) S
|
nr_node_for_each_safe(a,
- b,
c, d) S
|
- for_each_gfn_sp(a, c, d, b) S
+ for_each_gfn_sp(a, c, d) S
|
- for_each_gfn_indirect_valid_sp(a, c, d, b) S
+ for_each_gfn_indirect_valid_sp(a, c, d) S
|
for_each_host(a,
- b,
c) S
|
for_each_host_safe(a,
- b,
c, d) S
|
for_each_mesh_entry(a,
- b,
c, d) S
)
...+>
[akpm@linux-foundation.org: drop bogus change from net/ipv4/raw.c]
[akpm@linux-foundation.org: drop bogus hunk from net/ipv6/raw.c]
[akpm@linux-foundation.org: checkpatch fixes]
[akpm@linux-foundation.org: fix warnings]
[akpm@linux-foudnation.org: redo intrusive kvm changes]
Tested-by: Peter Senna Tschudin <peter.senna@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-28 09:06:00 +08:00
|
|
|
hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
|
2011-04-04 13:30:58 +08:00
|
|
|
common.hnode) {
|
|
|
|
qfq_destroy_class(sch, cl);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
qdisc_class_hash_destroy(&q->clhash);
|
|
|
|
}
|
|
|
|
|
|
|
|
static const struct Qdisc_class_ops qfq_class_ops = {
|
|
|
|
.change = qfq_change_class,
|
|
|
|
.delete = qfq_delete_class,
|
net_sched: remove tc class reference counting
For TC classes, their ->get() and ->put() are always paired, and the
reference counting is completely useless, because:
1) For class modification and dumping paths, we already hold RTNL lock,
so all of these ->get(),->change(),->put() are atomic.
2) For filter bindiing/unbinding, we use other reference counter than
this one, and they should have RTNL lock too.
3) For ->qlen_notify(), it is special because it is called on ->enqueue()
path, but we already hold qdisc tree lock there, and we hold this
tree lock when graft or delete the class too, so it should not be gone
or changed until we release the tree lock.
Therefore, this patch removes ->get() and ->put(), but:
1) Adds a new ->find() to find the pointer to a class by classid, no
refcnt.
2) Move the original class destroy upon the last refcnt into ->delete(),
right after releasing tree lock. This is fine because the class is
already removed from hash when holding the lock.
For those who also use ->put() as ->unbind(), just rename them to reflect
this change.
Cc: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2017-08-25 07:51:29 +08:00
|
|
|
.find = qfq_search_class,
|
2017-05-17 17:07:55 +08:00
|
|
|
.tcf_block = qfq_tcf_block,
|
2011-04-04 13:30:58 +08:00
|
|
|
.bind_tcf = qfq_bind_tcf,
|
|
|
|
.unbind_tcf = qfq_unbind_tcf,
|
|
|
|
.graft = qfq_graft_class,
|
|
|
|
.leaf = qfq_class_leaf,
|
|
|
|
.qlen_notify = qfq_qlen_notify,
|
|
|
|
.dump = qfq_dump_class,
|
|
|
|
.dump_stats = qfq_dump_class_stats,
|
|
|
|
.walk = qfq_walk,
|
|
|
|
};
|
|
|
|
|
|
|
|
static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
|
|
|
|
.cl_ops = &qfq_class_ops,
|
|
|
|
.id = "qfq",
|
|
|
|
.priv_size = sizeof(struct qfq_sched),
|
|
|
|
.enqueue = qfq_enqueue,
|
|
|
|
.dequeue = qfq_dequeue,
|
|
|
|
.peek = qdisc_peek_dequeued,
|
|
|
|
.init = qfq_init_qdisc,
|
|
|
|
.reset = qfq_reset_qdisc,
|
|
|
|
.destroy = qfq_destroy_qdisc,
|
|
|
|
.owner = THIS_MODULE,
|
|
|
|
};
|
|
|
|
|
|
|
|
static int __init qfq_init(void)
|
|
|
|
{
|
|
|
|
return register_qdisc(&qfq_qdisc_ops);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void __exit qfq_exit(void)
|
|
|
|
{
|
|
|
|
unregister_qdisc(&qfq_qdisc_ops);
|
|
|
|
}
|
|
|
|
|
|
|
|
module_init(qfq_init);
|
|
|
|
module_exit(qfq_exit);
|
|
|
|
MODULE_LICENSE("GPL");
|